Intuitive Plan Construction and Adaptive Plan Selection

Typical tasks of multi agent systems are effective coordination of single agents and their cooperation. Especially in dynamic environments, like the RoboCup soccer domain, the uncertainty of an opponent’s team behavior complicates coordinated team action. This paper presents a novel approach for intuitive multi agent plan construction and adaptive plan selection to attempt these tasks. We introduce a tool designed to represent plans like in tactical playbooks in human soccer which allows easy plan construction, editing and managing. Further we introduce a technique that provides adaptive plan selection in offensive situations by evaluating effectiveness of plans and their actions with statistically interpreted results to improve a team’s style of play. Using experts as a concept for abstracting information about a team’s interaction with another, makes fast accommodated plan selection possible. We briefly describe our software components, examine the performance of our implementation and give an example for rational plan selection in the RoboCup Small Size League. 1 Motivation and Related Work A common hypothesis is, that the more a team’s behavior is adapted to the opponent, the more effective it is. There are two major ways for adaption: One is to analyse the opponent and adapt the own team behavior to it (e.g. [2, 8]). The other way, introduced in this work, is to analyse the own team effectiveness corresponding to the opponent and adapt team behavior thereby. The idea behind our approach is to enhance the B-Smart software, developed for the RoboCup Small Size League, which uses more or less only reactive behavior selection, with a deliberative component providing strategic moves and multi agent coordination. Deliberative principles become more and more one of the most challenging aspects within the focal point of artificial intelligence in RoboCup (e.g. see team descriptions to appear in [7]). This approach is restricted to offensive game situations and uses predefined stepwise action sequences, like tactical playbooks in human soccer (cf. [9]), further called plans. The approach is based on the work of Bowling et al. [3, 4] who are using a play (multiagent plan)-based coordination and opponent-adaption for the CMDragons Small Size Team. However, our approach differs in the way of plan/play construction and adaptive selection. Our 1 For further information see: http://www.b-smart.de/ implementation is divided into two parts. One separate component for creating and managing plans, the Strategist, and one for plan selection, execution and assessment, the Coach.